distill — session knowledge → four-layer knowledge base

Turn "what this session just learned" into durable knowledge. This is the dual of foraging: foraging pulls external material into a project's library; distilling pushes internal experience out into persistent form.

Fill in paths and naming conventions for your own project and IDE. This document uses placeholder notation and is not tied to any specific toolchain.

Prerequisite: your agent already has (or you want to build) all four of these — persistent rules, persistent memory, skills, and decision records. If all you have is a single markdown notebook, this layering is over-engineering — just write it down.

1. Scan the session for distillable signals

Ask, category by category:

• What bug did you fix, and what was the root cause and the fix (especially anything counter-intuitive)?
• What gotcha did you hit (a tool's quirky behavior, an environment difference, an API that doesn't do what its docs say)?
• What decision got made, and were there real alternatives on the table?
• What pattern worked, or failed (at the methodology level)?
• What did the user correct ("that's wrong", "you missed X")? → the strongest signal — it means your default behavior systematically diverged from what they expected
• What did you hand-build a second time → a skill candidate

2. Assign each item to a layer

Scope first, layer second. For every candidate, ask: would this still hold in a different project?

• Still holds (a tool's quirks, how you collaborate, engineering discipline, reporting preferences) → global scope: the agent's global instruction file or global memory. Never write it into a project-scoped bucket — a project bucket is only read by sessions inside that project, so a genuinely portable lesson written there is forgotten the moment you switch projects.
• Only holds in this project (paths, accounts, deployment topology, business constraints, this repo's specific traps) → project scope, per the table below.
• Before writing a memory, confirm which bucket you're actually writing into. Memory buckets are usually derived from the working directory, so a renamed directory, or a path containing non-ASCII characters that collides on a hash, can silently route a memory into someone else's bucket.

Why this comes first: a full-archive scan once found that completely generic lessons — "verify before asserting," "concurrent-session discipline" — existed in only two of thirty-odd project buckets, and every later incident traced back to a project that lacked them. The lesson had been written down; it was just written in the wrong place. The same scan found one project's memory had silently landed inside a different project's library entirely, due to a path hash collision — undetected until then.

Memory splits into four kinds: user (who they are), feedback (how to collaborate with them — always with a why and a how), project (what's currently in flight), reference (a pointer to something external).

3. Core discipline

Three rules keep distillation from rotting:

1. Search before adding. Before adding a rule, memory, or skill, grep for what already exists. Already covered → update that one entry. Never create a duplicate.
2. Layer heuristic. An automatic guardrail → rule; a fact to recall → memory; a reusable workflow → skill; a decision with alternatives → decision record. The same lesson can legitimately be both a memory (the fact) and a rule (the guardrail) — but don't duplicate the content twice; the rule references the memory, it doesn't copy it.
3. Guard against bloat. Rule and skill descriptions are a standing cost — every one of them burns tokens on every single session. Only keep high-value ones; a description should say only "what it is + when to use it," never a pile of trigger keywords; reference, don't copy — point at what already exists instead of re-pasting it. When in doubt, don't add it — never pad for the sake of padding.

Why rule 1 comes first: in one real collaboration, the same mistake happened twice — once, a new document was created when the repository already had one covering the same ground, nobody had found it; the second time, an entire rename effort was carried out for a name that had already belonged to that account for two weeks, because a search had missed one namespace. Both traced back to the same root cause: nobody enumerated exhaustively before acting. A knowledge base rarely rots from having too little written down — it rots from the same fact being written three times, each version slightly contradicting the others, with nobody sure which one to trust.

4. Hygiene before landing anything

• Concurrent sessions: the repo may have another session's changes in flight → run git status first, and stage only your own specific paths (never -A or .) — don't sweep up someone else's in-progress work. Pull with --ff-only first if there's a remote.
• A new skill → also add whatever symlink or registration step your IDE needs, or it sits on disk invisible to the agent.
• A new memory → add one line to the memory file and one line to its index.
• A decision record with cascading follow-ups → verify every follow-up item actually landed where it was supposed to.
• A finished decision record → manually add a row to the ADR index's "recent" table. Freshness-checking tools usually don't validate this table, so a missing row drifts silently — "the tool passed" does not mean "the index is complete" (this exact row has been missed by three different authors in a row on one project).
• Run the project's checks (tests / lint / index checks) after any edit, and they must all pass before you commit — and run every check that pipeline has, not just the ones you're used to running. The one you skip is exactly the one that catches you in CI.
• Some IDE-managed directories are unreliable to write to via shell → use your editor tooling instead, and commit before you touch anything there.

5. Output

Give the user an allocation table: for each piece of knowledge, which scope it went into (global / project) × which layer (rule / memory / skill / decision record), why, and — just as important — what you decided not to add, and why (already covered / not high-value enough). Reporting what you left out is as important as reporting what you added — it's the proof you actually filtered, rather than dumping the whole session back out.

Don't

• Don't pad low-value memories or rules to look thorough. The bloat tax is real: every standing description gets billed on every future session.
• Don't commit another session's in-progress changes as if they were your own distillation.
• A big decision (an architecture change, a long-running initiative) → pair it with a decision record. A small operational detail → hang it off an existing workflow or rule instead. Don't open a decision record for every little thing.